Mechanistically Interpretable AI for Accelerated Energy Materials Design

Breakthroughs in energy materials stem from a systematic understanding of catalytic activity and stability at the atomic scale. However, the growing complexity of real-world energy applications, conflicting material characterization metrics, and the overwhelming volume of experimental data pose significant challenges in identifying fundamental structure-property relationships and translating them into transformative advancements. While materials informatics and data-driven approaches have accelerated discovery, their effectiveness is often hindered by dataset bias, limited interpretability, and poor generalizability. To address these challenges, we developed a Two-Stage Material Screening framework, integrating high-throughput computations, standardized experiments, and active learning to systematically explore a vast chemical space of 6,940,032 candidates, identifying 4,287 promising electrocatalysts. By leveraging SHAP-based analysis, we revealed the pivotal role of d-p band hybridization in oxygen reduction reaction electrocatalysis, effectively linking theoretical insights with experimental validation. Notably, protonic ceramic electrochemical cells incorporating five of the most promising electrocatalysts exhibited a record-breaking peak power density of 2.68 W cm ² at 600 °C – 35% higher than previous benchmarks – while maintaining exceptional durability over 500 hours. Our AI-driven approach accurately predicts material properties, reveals critical insights, and accelerates experimental validation, significantly advancing energy materials design.